skip to content

It is well documented that there are a number of instances in which the accuracy of empirical potentials falls short of our requirements – this compromise must often be accepted if we are to study systems any larger than a few hundred atoms. When greater accuracy is mandatory, we turn to ab initio molecular dynamics – almost always relying on density functional theory (DFT) for the determination of energies and forces. In such cases, however, we are often limited by the computational cost of these simulations to studying systems with no more than a few hundred atoms. Machine learning offers a way to perform simulations with the accuracy of ab initio methods, but with a cost which is much closer to that of traditional empirical potentials. We are interested in applying these machine learning potentials to understanding the behaviour of carbon, and the interface between water and carbonaceous systems. One particular case in which machine learning models may be useful is the case of water on graphene and hexagonal boron-nitride, two similar materials where error in the binding energy of just a few tens of meV can change their behaviour of from hydrophilic to hydrophobic. In this situation, machine learning potentials have the possibility to provide the accuracy which is required for reliable simulations, while doing so efficiently enough that long simulations with tens of thousands of atoms can be performed.

Related Publications 

An accurate and transferable machine learning potential for carbon.
P Rowe, VL Deringer, P Gasparotto, G Csányi, A Michaelides – The Journal of Chemical Physics (2020) 153, 034702
Machine Learning Potential for Hexagonal Boron Nitride Applied to Thermally and Mechanically Induced Rippling
FL Thiemann, P Rowe, EA Müller, A Michaelides – The Journal of Physical Chemistry C (2020) 124, 22278
Development of a machine learning potential for graphene
P Rowe, G Csányi, D Alfè, A Michaelides – Physical Review B (2018) 97, 054303